Arm announces AGI CPU with up to 136 cores for agent AI infrastructure in data centers

Arm introduced the Arm AGI CPU, its first production-ready proprietary silicon solution built on the Neoverse platform. The processor was developed to support the growing demand for agent AI infrastructure, where systems coordinate tasks autonomously and continuously on a global scale.

This initiative marks a significant expansion in the company’s offering, which now offers complete processors in addition to IP and subsystem licenses. The move addresses requests from partners seeking mature Arm platforms for rapid deployment in hyperscale data centers.

The rise of agent AI changes the central role of the CPU in modern infrastructure. Sistemas AI systems operate seamlessly, with software agents interacting with multiple models and executing decisions in real time without constant human intervention.

In this configuration, the CPU manages accelerator orchestration, memory allocation, storage, and data movement among thousands of distributed tasks. The new architecture seeks to eliminate bottlenecks that limited sustained performance in complex and parallel loads.

Rack-scale performance and density

The Arm AGI CPU is designed to deliver high performance per task under sustained load across thousands of simultaneous cores. Cada component, including frequency, memory subsystem, and I/O interfaces, is optimized for dense environments with power and cooling constraints typical of today’s data centers.

The reference configuration in a two-node 1U server uses two chips and totals 272 cores per blade. A standard 36 kW air-cooled rack accommodates 30 blades, resulting in 8160 cores.

In partnership with Supermicro, a 200 kW liquid-cooled design was developed that supports 336 CPUs and more than 45 thousand cores per rack. Nessas deployments, the solution delivers more than twice the performance per rack compared to recent equivalent x86 systems.

Several factors contribute to this advantage. Class-leading memory bandwidth enables more effective execution threads per rack, avoiding degradation seen in architectures competing for resources under heavy load.

Neoverse V3 cores provide better single-threaded processing, making each thread do more useful work. The combination of more usable threads and greater per-thread efficiency yields significant gains at the entire rack level.

• Up to 136 Neoverse V3 cores per CPU
• Memory bandwidth up to 6 GB/s per core with latency under 100 ns
• PCIe Gen6 support with 96 lanes and CXL 3.0
• TDP configurable in the range of 230 W to 420 W depending on the model

Partnerships and early adoption in the ecosystem

Meta acts as the main partner and customer in the joint development of the Arm AGI CPU. The processor has been optimized to integrate with the company’s gigawatt-scale infrastructure and work in conjunction with Meta’s custom MTIA accelerators.

Other launch partners include Cerebras, Cloudflare, F5, OpenAI, Positron, Rebellions, SAP, and SK Telecom. Cada organization collaborates on solution implementation to accelerate AI services in cloud, network and enterprise environments.

Leia Tambem

Colby Minifie confirms Ashley Barrett’s powers in The Boys season five

Napoli x Milan in Serie A with confirmed lineups and where to watch live

New battery test puts Galaxy S26 Ultra ahead of iPhone 17 Pro Max in global ranking

Commercial systems based on the Arm AGI CPU are now available for order from ASRockRack, Lenovo and Supermicro. Early availability allows providers to begin testing and deployment planning at scale.

Arm also released the Arm AGI CPU 1OU Dual Node reference server, in the Open Compute Project standard DC-MHS format. The company plans to contribute the open design, supporting firmware, architecture specifications, and diagnostic tools.

These contributions aim to accelerate adoption across the ecosystem and ensure broad compatibility with existing Arm-based software and hardware.

Expansion of the Arm platform to full silicon

The launch represents the beginning of a new line of silicon products for data centers from Arm. The AGI CPU is manufactured by TSMC on a 3nm process and uses high-performance cores with vector extensions for AI workloads.

Varied models include 136-core versions focused on overall performance, 128-core options optimized for cost, and 64-core variants prioritizing per-core memory bandwidth. Todas share Armv9.2 architecture with support for bfloat16 and INT8 instructions.

The initiative occurs in parallel to the Neoverse CSS computing subsystems roadmap, allowing customers to choose between custom IP, platform solutions or ready-made Arm processors.

More than 50 leading hyperscale, cloud, semiconductor, memory, networking and software companies support the expansion of the Arm platform to full silicon. Essa collaboration reinforces ecosystem maturity for agent AI workloads.

Next steps and availability

The Arm AGI CPU is now available for order, with volume production expected in the second half of 2026. The solution aims to provide an efficient computational foundation for continuous innovation in industries that rely on AI infrastructure.

Partners highlight benefits such as greater performance density, reduced energy consumption and better scalability for agent orchestration. The architecture enables precise coordination between CPUs and specialized accelerators in heterogeneous environments.

Arm remains committed to evolving the platform to meet future demands from fast-growing cloud providers and AI companies.